Text-to-speech (TTS)

A text-to-speech model takes text and a chosen voice and returns audio. Current models sound far more natural than older systems, and many can stream audio as it's generated, so playback starts before the whole clip is ready. Anthropic doesn't offer TTS, so Claude-based apps pair it with a speech provider.

TTS raises product questions plain text doesn't. Text written for the eye often sounds wrong spoken: tables, URLs, code, abbreviations, and long lists need rewriting for the ear, often by an LLM step before synthesis. Product names, numbers, and dates may be mispronounced, so you'll want pronunciation controls or substitutions. Latency matters too: stream the audio and split long text into sentences so the first words play quickly.

Pricing is usually by character or audio length, so caching audio for text that repeats, such as a fixed help-article narration, pays off. Custom and cloned voices need documented consent from the person whose voice it is, and you should tell listeners when a voice is AI-generated; some providers' usage policies require it, and laws are moving the same way. Treat the voice as a branding decision: one consistent voice across features makes them feel like one product.